Two standalone checks were written and left unwired, an hour apart, for the same correct reason

Both authors declined to add a fleet-wide gate step on the way past, which is the right discipline and produces a predictable failure: a good row sits unwired, waiting on a decision nobody knows they are holding, and gets re-proposed in a fortnight by someone who did not find the first.

One of the two has since landed — ef96b48f8 wired the HEAD-side lib/rtl sweep, armed off the merge-base, at 0s on a clone that did not touch compiler/. That sets the precedent and answers the pricing question for its shape. This row is the other half.

tools/lowering_passthrough_census.py is not a sketch. Its design point is worth reading before pricing it: the raw set difference is not the census. 83 AST kinds have a value arm and no address arm — nearly all statements and literals, so a guard on that flags everything and means nothing. Requiring the value arm to be a pass-through (Result := IRLowerAST(ASTLeft/Right[node]), a kind that forwards to a child and therefore has an address its parent could want) takes 83 to 4. Two controls, both branched on and both verified against the real pre-fix tree: --self-check must NAME AN_STR_FROM_CHAR, and a normal run over that tree must exit 1 with it marked NEW.

Its ACCEPTED list is deliberately not a clean bill: one entry was constructed against gcc and matches, two were not constructed either way, and the file says "reachability asked and not demonstrated" in those words — neither a defect nor a clearance. An exemption list that cannot distinguish checked and fine from not checked is a clean bill of health written by nobody.

The pricing question is not per-candidate. It is "how much may quick grow, and what buys the most", and it cannot be answered one row at a time — which is the reason this exists as a ticket rather than as a sentence in a commit message nobody will re-read.